ip_adapter/ip_adapter.py

import os
from typing import List

import torch
from diffusers import StableDiffusionPipeline
from diffusers.pipelines.controlnet import MultiControlNetModel
from PIL import Image
from safetensors import safe_open
from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
from transformers import Qwen2VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

from .utils import is_torch2_available, get_generator

if is_torch2_available():
    from .attention_processor import (
        AttnProcessor2_0 as AttnProcessor,
    )
    from .attention_processor import (
        CNAttnProcessor2_0 as CNAttnProcessor,
    )
    from .attention_processor import (
        IPAttnProcessor2_0 as IPAttnProcessor,
    )
    from .attention_processor_faceid import (
        LoRAAttnProcessor2_0 as LoRAAttnProcessor,
    )
    from .attention_processor_faceid import (
        LoRAIPAttnProcessor2_0 as LoRAIPAttnProcessor,
    )     
else:
    from .attention_processor import AttnProcessor, CNAttnProcessor, IPAttnProcessor
    from .attention_processor_faceid import LoRAAttnProcessor, LoRAIPAttnProcessor
from .resampler import Resampler


class ImageProjModel(torch.nn.Module):
    """Projection Model"""

    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
        super().__init__()

        self.generator = None
        self.cross_attention_dim = cross_attention_dim
        self.clip_extra_context_tokens = clip_extra_context_tokens
        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
        self.norm = torch.nn.LayerNorm(cross_attention_dim)

    def forward(self, image_embeds):
        embeds = image_embeds
        clip_extra_context_tokens = self.proj(embeds).reshape(
            -1, self.clip_extra_context_tokens, self.cross_attention_dim
        )
        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
        return clip_extra_context_tokens


class MLPProjModel(torch.nn.Module):
    """SD model with image prompt"""
    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024):
        super().__init__()
        
        self.proj = torch.nn.Sequential(
            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
            torch.nn.GELU(),
            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
            torch.nn.LayerNorm(cross_attention_dim)
        )
        
    def forward(self, image_embeds):
        clip_extra_context_tokens = self.proj(image_embeds)
        return clip_extra_context_tokens


class IPAdapter:
    def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_tokens=4):
        self.device = device
        self.image_encoder_path = image_encoder_path
        self.ip_ckpt = ip_ckpt
        self.num_tokens = num_tokens

        self.pipe = sd_pipe.to(self.device)
        self.set_ip_adapter()

        # load image encoder
        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
            self.device, dtype=torch.float16
        )
        self.clip_image_processor = CLIPImageProcessor()
        # image proj model
        self.image_proj_model = self.init_proj()

        self.load_ip_adapter()

    def init_proj(self):
        image_proj_model = ImageProjModel(
            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
            clip_embeddings_dim=self.image_encoder.config.projection_dim,
            clip_extra_context_tokens=self.num_tokens,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model

    def set_ip_adapter(self):
        unet = self.pipe.unet
        attn_procs = {}
        for name in unet.attn_processors.keys():
            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
            if name.startswith("mid_block"):
                hidden_size = unet.config.block_out_channels[-1]
            elif name.startswith("up_blocks"):
                block_id = int(name[len("up_blocks.")])
                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
            elif name.startswith("down_blocks"):
                block_id = int(name[len("down_blocks.")])
                hidden_size = unet.config.block_out_channels[block_id]
            if cross_attention_dim is None:
                attn_procs[name] = AttnProcessor()
            else:
                attn_procs[name] = IPAttnProcessor(
                    hidden_size=hidden_size,
                    cross_attention_dim=cross_attention_dim,
                    scale=1.0,
                    num_tokens=self.num_tokens,
                ).to(self.device, dtype=torch.float16)
        unet.set_attn_processor(attn_procs)
        if hasattr(self.pipe, "controlnet"):
            if isinstance(self.pipe.controlnet, MultiControlNetModel):
                for controlnet in self.pipe.controlnet.nets:
                    controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
            else:
                self.pipe.controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))

    def load_ip_adapter(self):
        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
            state_dict = {"image_proj": {}, "ip_adapter": {}}
            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
                for key in f.keys():
                    if key.startswith("image_proj."):
                        state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
                    elif key.startswith("ip_adapter."):
                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
        else:
            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
        self.image_proj_model.load_state_dict(state_dict["image_proj"])
        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
        ip_layers.load_state_dict(state_dict["ip_adapter"])

    @torch.inference_mode()
    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
        if pil_image is not None:
            if isinstance(pil_image, Image.Image):
                pil_image = [pil_image]
            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
        else:
            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
        uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
        return image_prompt_embeds, uncond_image_prompt_embeds

    def set_scale(self, scale):
        for attn_processor in self.pipe.unet.attn_processors.values():
            if isinstance(attn_processor, IPAttnProcessor) or isinstance(attn_processor, LoRAIPAttnProcessor):
                attn_processor.scale = scale

    def generate(
        self,
        pil_image=None,
        clip_image_embeds=None,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        guidance_scale=7.5,
        num_inference_steps=30,
        **kwargs,
    ):
        self.set_scale(scale)

        if pil_image is not None:
            num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
        else:
            num_prompts = clip_image_embeds.size(0)

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
            pil_image=pil_image, clip_image_embeds=clip_image_embeds
        )
        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)

        with torch.inference_mode():
            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
                prompt,
                device=self.device,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds_, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat([negative_prompt_embeds_, uncond_image_prompt_embeds], dim=1)

        generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            guidance_scale=guidance_scale,
            num_inference_steps=num_inference_steps,
            generator=generator,
            **kwargs,
        ).images

        return images


class IPAdapterXL(IPAdapter):
    """SDXL"""

    @torch.inference_mode()
    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
        if pil_image is not None:
            if isinstance(pil_image, Image.Image):
                pil_image = [pil_image]
            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
        else:
            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
        clip_image_embeds = clip_image_embeds.mean(0, keepdim=True)
        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
        # if image_prompt_embeds.shape[0] > 1:
        #     image_prompt_embeds = image_prompt_embeds.mean(0, keepdim=True)
        #     uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds[:1]))
        # else:
        uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
        return image_prompt_embeds, uncond_image_prompt_embeds

    def generate(
        self,
        pil_image,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        num_inference_steps=30,
        **kwargs,
    ):
        self.set_scale(scale)

        # num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
        num_prompts = 1

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image)
        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)

        with torch.inference_mode():
            (
                prompt_embeds,
                negative_prompt_embeds,
                pooled_prompt_embeds,
                negative_pooled_prompt_embeds,
            ) = self.pipe.encode_prompt(
                prompt,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)

        self.generator = get_generator(seed, self.device)
        
        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            num_inference_steps=num_inference_steps,
            generator=self.generator,
            **kwargs,
        ).images

        return images


class IPAdapterPlus(IPAdapter):
    """IP-Adapter with fine-grained features"""

    def init_proj(self):
        image_proj_model = Resampler(
            dim=self.pipe.unet.config.cross_attention_dim,
            depth=4,
            dim_head=64,
            heads=12,
            num_queries=self.num_tokens,
            embedding_dim=self.image_encoder.config.hidden_size,
            output_dim=self.pipe.unet.config.cross_attention_dim,
            ff_mult=4,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model

    @torch.inference_mode()
    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
        if isinstance(pil_image, Image.Image):
            pil_image = [pil_image]
        clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
        clip_image = clip_image.to(self.device, dtype=torch.float16)
        clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
        uncond_clip_image_embeds = self.image_encoder(
            torch.zeros_like(clip_image), output_hidden_states=True
        ).hidden_states[-2]
        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
        return image_prompt_embeds, uncond_image_prompt_embeds


class IPAdapterFull(IPAdapterPlus):
    """IP-Adapter with full features"""

    def init_proj(self):
        image_proj_model = MLPProjModel(
            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
            clip_embeddings_dim=self.image_encoder.config.hidden_size,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model


class IPAdapterPlusXL(IPAdapter):
    """SDXL"""

    def init_proj(self):
        image_proj_model = Resampler(
            dim=1280,
            depth=4,
            dim_head=64,
            heads=20,
            num_queries=self.num_tokens,
            embedding_dim=self.image_encoder.config.hidden_size,
            output_dim=self.pipe.unet.config.cross_attention_dim,
            ff_mult=4,
        ).to(self.device, dtype=torch.float16)
        return image_proj_model

    @torch.inference_mode()
    def get_image_embeds(self, pil_image):
        if isinstance(pil_image, Image.Image):
            pil_image = [pil_image]
        clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
        clip_image = clip_image.to(self.device, dtype=torch.float16)
        clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
        uncond_clip_image_embeds = self.image_encoder(
            torch.zeros_like(clip_image), output_hidden_states=True
        ).hidden_states[-2]
        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
        return image_prompt_embeds, uncond_image_prompt_embeds

    def generate(
        self,
        pil_image,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        num_inference_steps=30,
        **kwargs,
    ):
        self.set_scale(scale)

        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image)
        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)

        with torch.inference_mode():
            (
                prompt_embeds,
                negative_prompt_embeds,
                pooled_prompt_embeds,
                negative_pooled_prompt_embeds,
            ) = self.pipe.encode_prompt(
                prompt,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)

        generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            num_inference_steps=num_inference_steps,
            generator=generator,
            **kwargs,
        ).images

        return images


class EasyRef(IPAdapter):
    """EasyRef-SDXL"""

    def __init__(self, sd_pipe, multimodal_llm_path, ip_ckpt, device, num_tokens=64, use_lora=False, lora_rank=128, cond_image_size=336):
        self.device = device
        self.multimodal_llm_path = multimodal_llm_path
        self.ip_ckpt = ip_ckpt
        self.num_tokens = num_tokens
        self.use_lora = use_lora
        self.lora_rank = lora_rank

        self.pipe = sd_pipe.to(self.device)
        self.set_ip_adapter()

        # load image encoder
        mllm_final_layer = Qwen2VLForConditionalGeneration.from_pretrained(
            multimodal_llm_path, 
            torch_dtype=torch.bfloat16, 
            attn_implementation="sdpa",
            device_map="cuda"
        )
        mllm_final_layer = mllm_final_layer.model
        mllm_final_layer.layers = mllm_final_layer.layers[-1:]
        mllm_final_layer.embed_tokens = torch.nn.Identity()
        mllm_final_layer.visual = torch.nn.Identity()
        mllm_final_layer.lm_head = torch.nn.Identity()
        mllm_final_layer.reference_tokens = torch.nn.Parameter(0.1 * torch.randn(num_tokens, mllm_final_layer.config.hidden_size))
        self.mllm_final_layer = mllm_final_layer.to(self.device)
        for i in range(len(self.mllm_final_layer.layers)):
            self.mllm_final_layer.layers[i].self_attn.is_causal = False

        multimodal_llm = Qwen2VLForConditionalGeneration.from_pretrained(
            multimodal_llm_path, 
            torch_dtype=torch.bfloat16, 
            attn_implementation="sdpa",
            device_map="cuda"
        )
        multimodal_llm.model.layers = multimodal_llm.model.layers[:-1]
        multimodal_llm.norm = torch.nn.Identity()
        self.multimodal_llm = multimodal_llm.to(self.device)

        min_pixels = ((cond_image_size // 28 - 1)**2) * 28 * 28
        max_pixels = ((cond_image_size // 28)**2 + 1) * 28 * 28
        self.image_processor = AutoProcessor.from_pretrained(
            multimodal_llm_path, min_pixels=min_pixels, max_pixels=max_pixels)
        # image proj model
        self.image_proj_model = self.init_proj()

        self.load_ip_adapter()

    def load_ip_adapter(self):
        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
            state_dict = {"image_proj_model": {}, "mllm_final_layer": {}, "unet": {}, "multimodal_llm": {}}
            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
                for key in f.keys():
                    if key.startswith("image_proj_model."):
                        state_dict["image_proj_model"][key.replace("image_proj_model.", "")] = f.get_tensor(key)
                    elif key.startswith("mllm_final_layer."):
                        state_dict["mllm_final_layer"][key.replace("mllm_final_layer.", "")] = f.get_tensor(key) 
                    elif key.startswith("multimodal_llm."):
                        state_dict["multimodal_llm"][key.replace("multimodal_llm.", "")] = f.get_tensor(key)                                                
                    elif key.startswith("unet."):
                        state_dict["unet"][key.replace("unet.", "")] = f.get_tensor(key)
        else:
            state_dict = {"image_proj_model": {}, "mllm_final_layer": {}, "unet": {}, "multimodal_llm": {}}
            f = torch.load(self.ip_ckpt, map_location="cpu")["module"]
            for key in f.keys():
                if key.startswith("image_proj_model."):
                    state_dict["image_proj_model"][key.replace("image_proj_model.", "")] = f[key]
                elif key.startswith("mllm_final_layer."):
                    state_dict["mllm_final_layer"][key.replace("mllm_final_layer.", "")] = f[key]
                elif key.startswith("multimodal_llm."):
                    state_dict["multimodal_llm"][key.replace("multimodal_llm.", "")] = f[key]
                elif key.startswith("unet."):
                    state_dict["unet"][key.replace("unet.", "")] = f[key]
        if len(list(state_dict["multimodal_llm"].keys())) > 0:
            self.multimodal_llm.load_state_dict(state_dict["multimodal_llm"], strict=False)
        self.image_proj_model.load_state_dict(state_dict["image_proj_model"])
        self.mllm_final_layer.load_state_dict(state_dict["mllm_final_layer"])
        unet_state_dict = self.pipe.unet.state_dict()
        unet_state_dict.update(state_dict["unet"])
        self.pipe.unet.load_state_dict(unet_state_dict, strict=False)        
        # ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
        # ip_layers.load_state_dict(state_dict["ip_adapter"])
                
    def set_ip_adapter(self):
        unet = self.pipe.unet
        attn_procs = {}
        for name in unet.attn_processors.keys():
            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
            if name.startswith("mid_block"):
                hidden_size = unet.config.block_out_channels[-1]
            elif name.startswith("up_blocks"):
                block_id = int(name[len("up_blocks.")])
                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
            elif name.startswith("down_blocks"):
                block_id = int(name[len("down_blocks.")])
                hidden_size = unet.config.block_out_channels[block_id]
            if cross_attention_dim is None:
                if self.use_lora:
                    attn_procs[name] = LoRAAttnProcessor(hidden_size=hidden_size, cross_attention_dim=cross_attention_dim, rank=self.lora_rank).to(self.device, dtype=torch.float16)
                else:                    
                    attn_procs[name] = AttnProcessor().to(self.device, dtype=torch.float16)
            else:
                if self.use_lora:
                    attn_procs[name] = LoRAIPAttnProcessor(
                        hidden_size=hidden_size,
                        cross_attention_dim=cross_attention_dim,
                        scale=1.0,
                        num_tokens=self.num_tokens,
                        rank=self.lora_rank,
                    ).to(self.device, dtype=torch.float16)
                else:
                    attn_procs[name] = IPAttnProcessor(
                        hidden_size=hidden_size,
                        cross_attention_dim=cross_attention_dim,
                        scale=1.0,
                        num_tokens=self.num_tokens,
                    ).to(self.device, dtype=torch.float16)
        unet.set_attn_processor(attn_procs)
        if hasattr(self.pipe, "controlnet"):
            if isinstance(self.pipe.controlnet, MultiControlNetModel):
                for controlnet in self.pipe.controlnet.nets:
                    controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
            else:
                self.pipe.controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))

    def init_proj(self):
        image_proj_model = MLPProjModel(
            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
            clip_embeddings_dim=self.multimodal_llm.config.hidden_size,
        ).to(self.device, dtype=torch.bfloat16)
        return image_proj_model

    @torch.inference_mode()
    def get_image_embeds(self, pil_image, system_prompt):
        if isinstance(pil_image, Image.Image):
            pil_image = [pil_image]
        data = []
        messages = [
            {
                "role": "user",
                "content": [],
            }
        ]     
        for image in pil_image:
            messages[0]["content"].append({"type": "image", "image": image})
        messages[0]["content"].append({"type": "text", "text": system_prompt})
        prompt = self.image_processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )
        image_inputs, video_inputs = process_vision_info(messages)
        inputs = self.image_processor(
            text=[prompt],
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
        )
        data.append(inputs)
        input_ids = torch.stack([example["input_ids"] for example in data], dim=0).to(self.device)
        attention_mask = torch.cat([example["attention_mask"] for example in data], dim=0).to(self.device)
        pixel_values = [example["pixel_values"] for example in data]
        image_grid_thw = torch.stack([example["image_grid_thw"] for example in data], dim=0).to(self.device)

        with torch.no_grad():
            inputs_embeds = self.multimodal_llm.model.embed_tokens(input_ids)
            new_inputs_embeds = []
            for i in range(len(pixel_values)):
                pixel_value = pixel_values[i].type(self.multimodal_llm.visual.get_dtype()).to(inputs_embeds.device)
                grid_thw = image_grid_thw[i]
                cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
                    dim=0, dtype=torch.int32
                )
                cu_seqlens = torch.nn.functional.pad(cu_seqlens, (1, 0), value=0)  
                image_embeds = []          
                for j in range(1, len(cu_seqlens)):
                    image_embed = self.multimodal_llm.visual(pixel_value[cu_seqlens[j - 1] : cu_seqlens[j]], grid_thw=grid_thw[(j - 1) : j]).to(inputs_embeds.device)
                    image_embeds.append(image_embed)
                image_embeds = torch.cat(image_embeds, dim=0)
                image_mask = input_ids[i] == self.multimodal_llm.config.image_token_id
                inputs_embed = inputs_embeds[i].clone()
                inputs_embed[image_mask] = image_embeds
                new_inputs_embeds.append(inputs_embed)
            inputs_embeds = torch.cat(new_inputs_embeds, dim=0)
            image_embeds = self.multimodal_llm(
                attention_mask=attention_mask,
                inputs_embeds=inputs_embeds,
                output_hidden_states=True
            ).hidden_states[-1]

            reference_tokens = self.mllm_final_layer.reference_tokens.to(self.device)
            image_embeds = torch.cat([image_embeds, reference_tokens.unsqueeze(0).repeat(image_embeds.shape[0], 1, 1)], dim=1).to(dtype=torch.bfloat16)
            attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, :reference_tokens.shape[0]])], dim=1)
            outputs = self.mllm_final_layer(
                attention_mask=attention_mask.to(self.device),
                inputs_embeds=image_embeds.to(self.device),
                output_hidden_states=True,
            )
            image_embeds = outputs.hidden_states[-1]
            image_embeds_ = []
            for image_embed in image_embeds:
                new_image_embed = image_embed[-reference_tokens.shape[0]:]
                image_embeds_.append(new_image_embed)
            image_prompt_embeds = self.image_proj_model(torch.stack(image_embeds_)).to(dtype=torch.float16)     
        return image_prompt_embeds

    def generate(
        self,
        pil_image,
        system_prompt,
        prompt=None,
        negative_prompt=None,
        scale=1.0,
        num_samples=4,
        seed=None,
        num_inference_steps=30,
        **kwargs,
    ):
        self.set_scale(scale)

        # num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
        num_prompts = 1

        if prompt is None:
            prompt = "best quality, high quality"
        if negative_prompt is None:
            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"

        if not isinstance(prompt, List):
            prompt = [prompt] * num_prompts
        if not isinstance(negative_prompt, List):
            negative_prompt = [negative_prompt] * num_prompts

        # image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image)
        image_prompt_embeds = self.get_image_embeds(pil_image, system_prompt[0])
        uncond_image_prompt_embeds = self.get_image_embeds(Image.new(mode="RGB", size=(int(512), int(512))), system_prompt[1])
        bs_embed, seq_len, _ = image_prompt_embeds.shape
        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)

        with torch.inference_mode():
            (
                prompt_embeds,
                negative_prompt_embeds,
                pooled_prompt_embeds,
                negative_pooled_prompt_embeds,
            ) = self.pipe.encode_prompt(
                prompt,
                num_images_per_prompt=num_samples,
                do_classifier_free_guidance=True,
                negative_prompt=negative_prompt,
            )
            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)

        generator = get_generator(seed, self.device)

        images = self.pipe(
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            num_inference_steps=num_inference_steps,
            generator=generator,
            **kwargs,
        ).images

        return images